LSP-DETR (Local Star Polygon DEtection TRansformer) is a lightweight, efficient, and end-to-end deep learning model for nuclei instance segmentation in histopathological images. It combines a DETR-based transformer decoder with star-convex polygon shape descriptors to enable accurate and fast segmentation without complex post-processing.

NOTE: This checkpoint was pretrained on the complete PanNuke dataset without class labels. This configuration prioritizes maximum recall, making it ideal for robust nucleus localization.

python

from transformers import AutoModelForObjectDetection, AutoImageProcessor


processor = AutoImageProcessor.from_pretrained(
    "RationAI/LSP-DETR", trust_remote_code=True
)
model = AutoModelForObjectDetection.from_pretrained(
    "RationAI/LSP-DETR", trust_remote_code=True
)

inputs = processor(img, device=device, return_tensors="pt")
outputs = model(**inputs)
results = processor.post_process(outputs)
results = processor.post_process_instance(
  results, height=img.size[1], width=img.size[0]
)

Citing LSP-DETR

BibTeX

@misc{pekar2026lspdetr,
  title={LSP-DETR: Efficient and Scalable Nuclei Segmentation in Whole Slide Images}, 
  author={Matěj Pekár and Vít Musil and Rudolf Nenutil and Petr Holub and Tomáš Brázdil},
  year={2026},
  eprint={2601.03163},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2601.03163}
}